What the system does, and what it refuses to do.
Plain description of how a photo becomes a millimetre, how the ranking is computed, and exactly what this prototype does not do yet.
Widths come from marker-scaled geometry. The classifier decides the mechanism, never the number.
A single score is a snapshot. The ledger stores slopes, so a wall can be ranked by how fast it is failing.
Anything crossing the intervention line escalates to an engineer. The system triages, it does not certify.
Pipeline
A guardian photographs the face with a printed A4 fiducial marker in frame. WhatsApp or the web app; works offline and uploads later.
Marker detection solves scale; angle, sharpness and lighting are checked. Anything unusable is refused at intake — never guessed.
Segmentation localises damage and classical geometry returns width in millimetres. A colour-and-texture screen reports salt crust, damp and surface loss as area percentages.
Each accepted capture is compared to the structure’s history. Least-squares slope gives mm per month, which is what predicts failure.
Rate, exposure, salt loading and structural criticality produce a risk score; a budget-constrained solver turns the portfolio into an ordered works plan.
A work order with a traditional-material bill of quantities is issued, and the next capture is scheduled. The ledger grows.
Failure physics we score against
Diagnostic work on historic buildings in Najran identifies clay-mineral dissolution and salt crystallisation as the dominant degradation drivers, accelerated by wetting–drying cycles above 44 °C, wind erosion and water entering the mud cavity. Mean compressive strength of the earthen wall samples is 2 MPa, with water absorption at the 15% ceiling.
That is why the model scores efflorescence, moisture at the rising damp line and facade orientation alongside crack width — a pure crack detector would be blind to the mechanism that actually kills the wall.
Abdullah, Abd El-Aal, Radwan, Al-Awah. Numerical simulation and diagnosis geotechnical parameters of historical buildings in Najran City. Scientific Reports 13, 16968 (2023). DOI 10.1038/s41598-023-43959-1
Model card — honest version
What is mocked in this prototype
- Marker-scaled measurement: ArUco scale solve, medial-axis crack width in millimetres
- Multiple defects per capture, each with its own width and length; the widest sets the risk band
- Surface screen for salt crust, damp and surface loss — colour and texture statistics, not a trained model
- Refusal gates — no marker, or no coherent damage, returns no number at all
- Risk scoring, rates, thresholds and the budget-constrained ranking
- Work order generation, bill of quantities and cost roll-up
- Ledger persistence, filtering, sorting and the record per structure
- Trained mechanism classification: the surface screen is a heuristic, not a model trained on labelled Najran walls — it reports indices and a dominant reading, with no confusion matrix behind it
- Salt loading and moisture are not measured — they read as zero on a real capture
- Structures, guardians, villages and costs: synthetic catalogue
- Authentication: role selection only, no credentials